In order to mitigate the effects of climate change, it is important to maximize the power output of our wind energy resources. Wind farm control is typically greedy, with all turbines facing directly into the wind. Due to wake interactions, this leads to reductions in power for downwind turbines for some wind conditions. An alternative is wake steering, which involves misaligning the yaws of upwind turbines to deflect their wakes away from downwind turbines. This reduces the power generated by the upwind turbines but increases the power output of the farm as a whole. Tools such as FLORIS exist to model wakes and compute optimal wake steering yaw configurations for a given wind farm and wind conditions, but their runtime scales quadratically in the number of turbines, which prevents real-time control for large farms.
Using a graph neural network with wind turbines as nodes and wake interactions between turbines as edges, we learn to approximate the FLORIS model. Our trained model is able to find yaw configurations that outperform the greedy solution and often match or outperform FLORIS in linear time for wind farms with the same number of turbines as it was trained on. Relational inductive biases in graph neural networks also enable generalization to unseen wind farms, including those with different numbers of turbines. Our trained model is able to accurately estimate the power output of wind farms with ten times the number of turbines it was trained on, as well as find yaw configurations for these examples that are on average of higher quality than the greedy solution in linear time.
Matthew Travers (Advisor)
Zoom Participation. See announcement.